LLM agents automate HPC scaling, job translation, and bio workflows

A new paper from Vanessa Sochat and Daniel Milroy explores using LLM-powered agentic frameworks to automate high-performance computing (HPC) tasks. The framework combines a large language model with tool functions and resources to execute descriptive prose-like instructions. The authors tested it on three complex scenarios: optimizing and running a scaling study with low-latency networking on AWS, translating job specifications between Slurm and Flux workload managers, and designing an end-to-end biosciences workflow. Results show the agents successfully completed all tasks but surfaced task-specific failure modes.

In the scaling study, agents deployed and optimized applications but monitored running jobs inefficiently, preferring conservative fixed waits over event-driven subscriptions. For job translation, they converted specifications between Slurm and Flux with high accuracy, though processor-affinity flags were the most common error. Most impressively, the agent reproduced an expert-written variant-calling pipeline almost exactly, agreeing with the reference call set in 18 of 19 completed runs. The agent reached this result through many distinct yet functionally equivalent workflow implementations. The authors note this information is invaluable for developing multi-cluster setups with scheduling and transformation handled by agents.